CALIBRATION OF SPECTROSCOPIC SENSORS WITH GAUSSIAN PROCESS AND VARIABLE SELECTION

Abstract Multivariate spectroscopic calibration models can be improved by selecting a subset of spectral variables that are identified as being the most informative in terms of inferring sample properties. This paper proposes the application of Gaussian processes for both variable selection and the development of calibration models. A Gaussian process is a Bayesian regression technique that assigns Gaussian priors over the regression functions. The covariance function of the Gaussian process is characterized by a number of hyper-parameters and by associating each spectral variable with a hyper-parameter, the relevance of the corresponding variable to the prediction can be automatically determined. Prior to the training of a Gaussian process using a Markov chain Monte Carlo approach, a pre-processing step is proposed based on a statistical significance test to reduce the computational time materialising from the large number of variables present within spectroscopic data. The methodology presented is applied to two sets of near infrared spectral data, and enhanced prediction performance is achieved when both the pre-processing step and a Gaussian process are implemented.